A public institute proved the hard part; the scientists who left with it now sell the state what a laboratory cannot build

The product

Proxima Fusion is building a fusion power plant in Bavaria. Its machine is a stellarator – a ring of plasma held by magnets twisted into a shape no one would draw by hand – rather than the tokamak most of the industry works on. The company was founded in April 2023 in Munich as the first spin-out in the sixty-year history of the Max Planck Institute for Plasma Physics, the institute that built and runs Wendelstein 7-X, and it now has around 200 scientists, engineers and operators in Munich, Zurich and Oxford.

Nothing is for sale. There are two machines ahead of it. Alpha, to be built next to the institute in Garching and operating in the early 2030s, is meant to be the first stellarator whose plasma releases more fusion power than the heating systems put into it. Stellaris, the power plant, is planned for Gundremmingen, on the site of a fission plant RWE is decommissioning, with electricity on the grid later in the 2030s.

How it works

The design came before the hardware. Proxima's first year went into StarFinder, a framework that connects plasma simulation, parametric CAD, magnets, support structures, divertor and vacuum vessel into a single loop an engineer can turn. In February 2025 the company published what came out of it: Stellaris, a power-plant concept written up with the institute, Karlsruhe, Lisbon and Wisconsin-Madison and peer-reviewed in Fusion Engineering and Design. On Sciortino's own slides it is a 3 GW-thermal, 1000 MW-electric machine with magnets above 20 tesla, and the reason those magnets matter is arithmetic: fusion power goes as the fourth power of the field, so moving from Wendelstein 7-X's 2.5 tesla to Stellaris's 9 gives about 170 times the power from a smaller device.

Then the metal. The Stellarator Model Coil, due in 2027, is a single non-planar high-temperature superconducting magnet storing more than 100 megajoules – the piece that decides whether the design can be manufactured at all. Around it the company has been assembling the things a magnet needs and Europe does not have: a framework with the Paul Scherrer Institute in Switzerland, its own cable line, a magnet factory planned at Gundremmingen, and since September 2026 a memorandum with Lower Saxony for the continent's first large-scale plant for the superconducting tape itself, which today is made at industrial scale mainly in Asia.

Some of the design work is simply given away. VMEC++, the company's rewrite of the Fortran equilibrium code every stellarator group depends on, went out under the MIT licence in January 2025 along with its internal documentation; six months later Proxima put three of its design problems, a dataset of about 160,000 plasma boundaries and a leaderboard on Hugging Face and asked the machine-learning world to beat its own baselines.

Business model

No customer pays Proxima for anything, and none will until a plant reaches the grid. The company runs on equity and public money – more than €650 million ($740 million) in under three years, of which €95 million is public grants – and the thing it sells today is a place in a project that governments want to exist.

The shape of a plant's finance is already sketched. Alpha is costed at about €2 billion. Under the February 2026 memorandum Proxima expects to put in roughly 20 per cent privately, Bavaria has indicated a state contribution of about the same, and the four partners are pooling their bids for federal money under Germany's High-Tech Agenda; Sciortino has written that Proxima's share is €400 million against Bavaria's €400 million. The supply chain is financed the same way: the tape plant's first two phases are €140 million, split evenly between public and private money, with Lower Saxony putting up €21 million and Proxima signing as anchor customer.

That leaves the valuation itself doing the work a price list normally does. Asked how an investor finances something at the edge of the possible, Sciortino's answer was blunt: "As long as the valuation continues to grow strongly and there is a multiplier ahead of us, it is financeable. VCs only get scared when there is no further value growth."

The spike

Francesco Sciortino spent his career on tokamaks – a PhD at MIT, then the institute in Garching, where he was one of Europe's coordinators for tokamak research. Then 2022 happened twice over. Wendelstein 7-X, built with more than €1.3 billion of German and European public money, hit its design targets, showing both that the physics held and that a shape that complicated could actually be manufactured to tolerance. In the same year Landreman and Paul showed that stellarator plasmas could be numerically optimised to precise symmetries, and the work extending that to quasi-isodynamic shapes followed a year later.

He left. "In 2022, we realized that tokamaks cannot be scaled to power plants," he told Munich Startup at the startup hub of the Munich Security Conference. "At the same time, it became clear that stellarators can be optimized numerically – further than I myself had ever thought possible." Four others founded the company with him, and in April 2023 the institute had its first spin-out in sixty years.

His reason is about physics. What they carried out of the institute was something else: a result a state had already paid €1.3 billion to establish, a public dataset going back decades, and the codes – one founder had spent his doctorate inside VMEC's numerics, working on Wendelstein 7-X data. None of it made the institute able to build a power plant, and Proxima has been plain about which side of that line it sits on. Full-scale prototypes, it wrote in 2025, "are too expensive and slow to build just for iterative learning. In 2022, our founding team saw that this wasn't only a limitation – it was a once-in-a-generation opportunity for disruption."

So the first thing they shipped was a design rather than a device, and they had strangers check it. That is what the money answered: when Cherry Ventures and Balderton led €130 million in June 2025, Plural's Ian Hogarth gave the reason plainly – "their peer-reviewed stellarator power plant design confirms fusion really can be commercially viable." The company had nothing else to show.

Then it stopped designing and started ordering. The September 2025 extension was headlined as hardware execution and spent on tape, support plates, a cable line, dummy coils and vacuum vessel sections; by February 2026 more than thirty industrial firms had signed up behind one machine, and the state, the utility and the institute had signed a plan to put Alpha in Garching and the plant at Gundremmingen.

Why it matters

Anyone who wanted fusion in Europe in 2023 had three options, and all of them asked the buyer to wait. The public route ran through Wendelstein 7-X as an experiment and ITER and DEMO as the eventual plant, on a schedule measured in decades. The private money had gone to tokamak companies, mostly American, whose magnet work Proxima's own slides credit as an inspiration. And a handful of other stellarator efforts existed, none of them with a plant design anyone outside had checked.

What none of them handed a government or a utility was a document you could sign against: an integrated design that met the physics and the engineering constraints at the same time, with the trade-offs published rather than asserted. One slide in Sciortino's deck is a table of thirteen physics and technology criteria against every previous class of stellarator work – papers, design-point studies, isolated technology studies – with the crosses left in; only Stellaris ticks the row. That is the step Proxima closed, and the reason a peer-reviewed paper functioned as a product.

The proof, for a company with nothing to measure in users, is who has since committed. In November 2023 a Hacker News thread about the startup was mostly people doing arithmetic on its €7 million pre-seed and concluding it was "peanuts", with one commenter writing that he was "always skeptical of the viability of fusion startups". Since then: €200 million by September 2025, more than €650 million by July 2026, and a €2.4 billion valuation in the largest fusion round ever raised in Europe. Bavaria has pledged up to €400 million of state money to a company rather than to a laboratory. RWE, a utility with a fission site to decommission, put in both money and the site. Google made its first investment in a European fusion company. More than fifty industrial suppliers have organised themselves around a machine that does not exist yet, and a second German state has agreed to co-fund a factory for a material Proxima has only promised to buy.

None of that is electricity. No stellarator anywhere has produced net fusion power, and the company is unusually careful about the difference: its own post on Q>1 insists that Alpha's target is scientific gain in the plasma, not the whole plant, and that phrases like "more energy than it consumes" belong to Stellaris and the 2030s. What has been demonstrated so far is that a design, checked in public, can move a state.

The lesson

The asset was never an invention. Germany and the EU had spent more than €1.3 billion proving a stellarator could be built, and in six decades the institute that owned that proof had never turned any of it into a company. Its founders left with the result rather than with an idea, and supplied the half a laboratory cannot do: engineering to a date, procurement, factories, a balance sheet that carries a €2 billion project. The institute did not become a rival – it leads the physics on Alpha, and its foundation is on the cap table. The rule: when a public body has proved something at great cost and has no way to sell it, start the business that supplies what it structurally lacks, and its sponsor becomes your backer rather than your competitor. The test: which expensive result in your field was proved by an organisation that cannot build or sell it, and what would it need from you?

Then there is what it buys. The tape its magnets are wound from is made at scale only far away, in Asia; Proxima did not go shopping but signed with a German state to have a plant built, put its name down as first customer, and let a government that wants the industry pay half. An order large enough – 20,000 kilometres for one machine, 40,000 for the next – is itself the argument for a second source.

Both put you on someone else's timetable: elections and budget rounds, or volumes you have not yet needed. The second test: which input comes from one distant supplier, and is five years of your demand enough to make somebody build it closer?